Although my previous flock() based solution is preventing udev scan to
get pre-mature super blocks, it in fact masks the root problem:
There is a stray close() on writeable fd.
Commit b2a1be83b85f ("btrfs-progs: mkfs: keep file descriptors open
during whole time") tries to solve the problem by extending the lifespan
of writeable fds, so that when the fds are closed, the fs is ensured to
be properly populated.
The problem is, that patch is not covering all cases, there is a stray
fd close just under our noses: open_ctree_fs_info().
The function would open the initial device, then use that initial fd to
open the btrfs, then we immediately close the initial fd, as later IO
would all go with the device fd.
That close() call is causing problem, especially for mkfs, as at that
stage the fs is still using a temporary super block, not using the valid
btrfs super magic number.
Thus udev scan would race with mkfs, and if udev wins the race, it would
get the temporary super block, making libblk not to detect the new
btrfs.
This patchset would address the problem by:
Make sure open_ctree*() calls all have corresponding close_ctree()
The first patch, as later we will only close the initial fd caused by
open_ctree_fs_info() during close_ctree().
Save the initial fd into btrfs_fs_info for open_ctree_fs_info()
And later close the initial fd during close_ctree().
Make sure open_ctree_fd() callers to properly close the fd
Just an extra cleanup.
This patchset would work even without the usage of flock() to block udev
scan, and since without flock() calls, the procedure is much simpler.
In theory, with this patchset, we can remove the flock(), as the flock() solution still has its own problem, as it can do infinite long wait.
kdave
commented
4 months ago
Although my previous flock() based solution is preventing udev scan to get pre-mature super blocks, it in fact masks the root problem:
There is a stray close() on writeable fd.
Commit b2a1be83b85f ("btrfs-progs: mkfs: keep file descriptors open during whole time") tries to solve the problem by extending the lifespan of writeable fds, so that when the fds are closed, the fs is ensured to be properly populated.
The problem is, that patch is not covering all cases, there is a stray fd close just under our noses: open_ctree_fs_info().
The function would open the initial device, then use that initial fd to open the btrfs, then we immediately close the initial fd, as later IO would all go with the device fd.
That close() call is causing problem, especially for mkfs, as at that stage the fs is still using a temporary super block, not using the valid btrfs super magic number.
Thus udev scan would race with mkfs, and if udev wins the race, it would get the temporary super block, making libblk not to detect the new btrfs.
This patchset would address the problem by:
Make sure open_ctree*() calls all have corresponding close_ctree() The first patch, as later we will only close the initial fd caused by open_ctree_fs_info() during close_ctree().
Save the initial fd into btrfs_fs_info for open_ctree_fs_info() And later close the initial fd during close_ctree().
Make sure open_ctree_fd() callers to properly close the fd Just an extra cleanup.
This patchset would work even without the usage of flock() to block udev scan, and since without flock() calls, the procedure is much simpler.
In theory, with this patchset, we can remove the flock(), as the flock() solution still has its own problem, as it can do infinite long wait.